Spatiotemporal Anomaly Detection in Distributed Acoustic Sensing Using a GraphDiffusion Model.

Area of Science:

• Geophysics
• Data Science
• Infrastructure Monitoring

Background:

• Distributed acoustic sensing (DAS) is vital for large-scale infrastructure monitoring, providing dense spatial and temporal data.
• Anomaly detection in DAS data is challenging due to spatial correlations and nonlinear temporal dynamics.
• Current methods often ignore sensor layout, treating data as images or sequences, losing spatial topology.

Purpose of the Study:

• To introduce GraphDiffusion, a novel generative anomaly detection model for DAS data.
• To explicitly model the spatial arrangement of DAS sensors and capture interchannel dependencies.
• To improve anomaly detection accuracy by preserving spatial and temporal information.

Main Methods:

• Developed GraphDiffusion, combining a conditional denoising diffusion probabilistic model (DDPM) and a graph neural network (GNN).
• Represented DAS channels as graph nodes with edges based on Euclidean proximity to model spatial layout.
• Utilized iterative denoising in the conditional DDPM to learn temporal dynamics of normal signals.

Main Results:

• GraphDiffusion achieved 98.2% F1K-AUC and 98.0% ROCK-AUC on real-world DAS datasets.
• The model outperformed comparative anomaly detection methods.
• Explicitly modeling spatial topology improved the detection of anomalies.

Conclusions:

• GraphDiffusion effectively detects anomalies in DAS data by integrating spatial and temporal information.
• The GNN-DDPM approach overcomes limitations of existing methods that disregard sensor layout.
• This model offers a significant advancement for reliable infrastructure monitoring using DAS.